ISCro4 Activity Improvement
In this project we asked two questions: whether the recombination activity of ISCro4 can be further improved, and whether an AI scientific agent can enable a team without formal training to carry out a complete directed-evolution campaign. Our results answer both affirmatively. Starting from the previously reported ISCro4(S30T/P54Q/S243H) variant, we screened ten DMS-nominated single mutations and identified the A225S variant, which exhibited a larger ΔCq than the parent, indicating measurably improved recombination activity. Notably, because the parent construct was already an engineered, activity-enhanced variant, the further gains we observed suggest that the existing DMS dataset still contains untapped beneficial mutations and that the activity ceiling of ISCro4 has not yet been reached.
Screening Assay
A key methodological lesson of this work is that the screening assay, more than the mutagenesis itself, often determines the success of a directed-evolution campaign. Our initial chloramphenicol-selection scheme failed because the terminator upstream of the resistance gene was insufficiently strong, allowing basal expression and chloramphenicol tolerance even without ISCro4-mediated recombination — a common pitfall of selection-based screens. Replacing genetic selection with a qPCR-based deletion assay gave us a quantitative, sensitive readout capable of resolving modest activity differences between variants, which is exactly what early rounds of evolution require. This two-plasmid qPCR format is low-cost, requires no specialized equipment beyond a qPCR instrument, and should be generalizable to other recombinases and transposon-associated systems. For future larger-library screens, selection stringency could be restored by using tandem terminators, a stronger insulated terminator, or a fluorescence-based readout.
AI Assistance and Experimental Verification
The most striking aspect of the project was the role of the AI agent. Biomni's contribution was not merely retrieving protocols: when plasmid construction repeatedly failed, it identified considerations — promoter properties, potential construct toxicity, and compatibility of genetic elements — that had occurred neither to us nor to our supervising teacher. In other words, the agent's value lay in connecting dispersed domain knowledge and surfacing the right questions, including questions we did not know to ask. Equally important, the agent's outputs functioned as hypotheses rather than answers: we verified every suggestion through literature review and experiment. We believe this "AI proposes, human verifies" loop is the appropriate model for student use of AI agents in research.
Study Limitations
Several limitations of our study should be acknowledged. First, activity was measured as plasmid-borne DNA deletion in E. coli; integration efficiency at genomic loci, maximum payload size, and editing fidelity (off-target rates) remain to be tested. Second, we evaluated single mutations only; combining the top substitutions may yield additive gains, but epistasis could also erase them. Third, our library was small — ten DMS-nominated sites — so the ISCro4 fitness landscape remains largely unexplored. Fourth, we do not yet have a structural or mechanistic explanation for why the improved mutations work. Finally, because our starting point was the engineered S30T/P54Q/S243H variant rather than wild-type ISCro4, improvements should be reported relative to that parent to avoid overstating the advance.
Next Steps
These limitations define our next steps. We plan to combine the most beneficial mutations and, if necessary, perform additional rounds of evolution using error-prone PCR; to model the improved variants structurally (e.g., with AlphaFold-based analysis) to rationalize their effects; to test integration of large donor payloads at genomic target sites; and to evaluate activity and specificity in mammalian cells, where ISCro4-based tools would have the greatest therapeutic and biotechnological relevance.
Accessibility and Human Oversight
Beyond the specific variants we obtained, this project illustrates a broader point. Directed evolution has historically been concentrated in well-resourced professional laboratories because its knowledge barrier is as limiting as its equipment barrier. Our experience suggests that AI agents can redistribute this "admission ticket" — not by replacing experimental work, but by compressing the knowledge gap between beginners and expert practice. At the same time, AI agents can make mistakes, and human oversight, experimental validation, and biosafety review remain indispensable. Used in this way, however, AI mentors can ensure that curiosity is no longer limited by experience or resources.
